The 52nd G7 leaders’ summit closed Wednesday afternoon in Évian and the final communique dropped late evening Paris time. The AI section is the part everyone was waiting for and the part most carefully written to commit to as little as possible. The seven nations affirm the importance of frontier AI safety, endorse the Hiroshima Process reporting framework as the international baseline, and agree to propose a more structured evaluations-and-incidents mechanism within 18 months. That is the actual sentence. The mechanism is not in the communique. The 18 months is.
This is approximately the fourth time since the original Hiroshima AI Code of Conduct in October 2023 that G7 leaders have agreed to agree on something later. The pattern has its own gravity now. Hiroshima 2023 produced the voluntary code. Italy 2024 produced an action plan to implement the code. Canada 2025 produced a refreshed action plan. Évian 2026 produces a deadline to propose a mechanism that the previous three documents implied was already coming. The 18-month clock now runs to roughly December 2027, which means the next G7 presidency, currently expected to be Italy again, will inherit it.
What is not in the communique is louder than what is. There is no language about model-weight access for sovereign safety review, which was the explicit Macron and Meloni ask going in. There is no pre-deployment notification commitment, which the European Commission floated in May as a minimum acceptable outcome. There is no reference to the Anthropic Fable 5 situation, the elephant that was actually in the room. There is one paragraph about export controls, the language of which the State Department reportedly negotiated down from “concern” to “noted recent developments.” The US delegation did its job.
The Europeans got the photograph of Altman, Amodei, and Hassabis sitting in a room with seven heads of state, which is a real political asset that did not exist last week and now exists. They did not get any binding instrument. Whether the photograph is worth more than the absent instrument depends on whether the photograph can be used to extract one in 18 months. The over-under on that is roughly even, which for international AI governance counts as a good day.